Wan 3.0 Release Date: Public Beta Is Live

The Wan 3.0 release date is no longer a question. Alibaba’s Tongyi Lab announced public beta on August 6, 2026, and the model is live now under the identifier wan3.0-video. Three headline capabilities shipped with it: native 30-second video generation, what Alibaba calls reality-grade rendering, and an omni-reference input system that accepts documents, spreadsheets, and slide decks alongside the usual text, image, audio, and video.

Pricing is published. Access is open. The Wan 3.0 release date turned out to be earlier than most coverage predicted. Most of the pages that spent months predicting the Wan 3.0 release date got the specifications wrong anyway.

What Shipped on the Wan 3.0 Release Date

Alibaba’s announcement on the Wan 3.0 release date lists three pillars:

Native 30-second video generation. Single-pass, no stitching. That doubles the 15-second ceiling of Wan 2.6 and Wan 2.7.

Reality-grade rendering. Alibaba’s framing is that every frame is built to look and feel more real, with more expressive characters, more consistent references, and better rendering of digital content.

Omni-reference. This is the genuinely new one. Beyond text, images, audio, and video, Wan 3.0 accepts structured documents as creative references. Supported input formats are doc, xls, ppt, pdf, txt, key, and pages. You can hand it a deck or a spreadsheet and it reads and creates from the contents.

There is a fourth change that Alibaba does not headline but that matters more for anyone building a workflow: Wan 3.0 is a single all-in-one model. Wan 2.7 split text-to-video, image-to-video, reference generation, and editing across four separate models. Wan 3.0 unifies reference, editing, replication, and motion driving into one, which removes the model-switching step from multi-stage pipelines.

Where to Access It

On the Wan 3.0 release date, public beta went live simultaneously on two platforms:

• Alibaba Cloud Model Studio — international site, ap-southeast-1 region, model ID wan3.0-video

• Qwen Cloud — listed at qwencloud.com under the same model name

A third route, wan.video, is described as coming very soon and will be members-only. Third-party platforms are still integrating; jxp.com’s own Wan 3.0 page is live with specifications but the generator is marked Coming Soon, which is the normal lag between an Alibaba beta and downstream API access.

Alibaba’s own wording is that full API access is rolling out soon, so beta access and general API availability are not the same thing yet. If you are planning an integration, treat the current window as evaluation rather than production.

Wan 3.0 Pricing

Alibaba published per-second rates in USD on the Wan 3.0 release date:

The 30-second column is arithmetic, not an official figure, but it is the number that matters if you are budgeting around the headline feature. A full-length 1080p generation costs roughly six dollars.

Note what is absent from that table: there is no 4K tier. The published ceiling is 1080p.

The 4K Claim Was Wrong, and Here Is Where It Came From

For months, “native 4K video” appeared on nearly every page speculating about the Wan 3.0 release date, and it survived right up to the Wan 3.0 release date itself. The launch pricing confirms it was never real. The confusion has a traceable origin.

Alibaba’s Model Studio documentation does describe a 4K capability in the Wan 2.7 family: wan2.7-image-pro supports resolutions up to 4096×4096, but only for text-to-image tasks where no image is input and image-set generation is disabled. Every other scenario caps at 2K. The API reference repeats the restriction in its own code comments.

Watch the claim mutate:

wan2.7-image-pro supports 4K for text-to-image ↓ “Wan 2.7 supports 4K” ↓ “Wan supports 4K” ↓ “Wan 3.0 will generate 4K video”

Each step drops a qualifier. By the fourth, the claim has changed model, changed modality, and changed version. The first statement is true and documented. The last was never confirmed by anyone at Alibaba, and the shipped pricing table settles it.

Which Wan 3.0 Release Date Predictions Were Right

Before launch, at least six incompatible Wan 3.0 release date predictions circulated. Now that the real Wan 3.0 release date is known, they can be scored:

The instructive result: the only source that got the Wan 3.0 release date right was the one that explicitly labelled its own date a rumor. The pages that stated dates with confidence were the pages that missed.

That pattern is worth remembering the next time a Wan 3.0 release date style question comes around for another model. Confidence in AI release coverage is inversely correlated with accuracy, because hedged speculation ranks worse than assertive speculation, so the hedges get stripped as claims propagate.

See what Wan 3.0 will offer on our Wan 3.0 page

Wan 3.0 vs Wan 2.7: What Actually Changed

The resolution row is the one to internalise, and it is where most pre-launch coverage of the Wan 3.0 release date went wrong. Wan 3.0 did not raise the resolution ceiling at all. The upgrade is duration, input flexibility, and consolidation — not pixels.

The Controls You Will Actually Use

Beyond the headline specifications, several practical behaviours shape how Wan 3.0 is prompted. These matter more day to day than duration numbers.

No negative prompt field. Wan 3.0 does not expose a traditional negative prompt input. Exclusions go into the main prompt as plain instructions — “no text on screen”, “do not change the character’s clothing”. If you are porting a prompt library from another model, this is the first thing to rewrite.

Stage your prompts past ten seconds. For clips longer than roughly ten seconds, describing the action in ordered stages produces more coherent results than a single dense paragraph. Give each stage a visible goal and state how the scene should end.

First-and-last-frame control, with a caveat. You can define opening and closing frames when a scene needs to move between two defined visual states. This control may become unavailable when multi-image reference mode is selected, so the two features are not always usable together.

Aspect ratio is either automatic or locked. The model can infer a suitable vertical, square, or landscape frame from the prompt, or you can lock output to 16:9, 9:16, or 1:1.

Audio is described, not uploaded. Synchronized audio is generated from prompt language. Dialogue should be written as the exact line you want spoken; ambient sound and action cues need naming explicitly.

Omni-Reference Is the Real Story

Native 30 seconds got the headline on the Wan 3.0 release date, but duration is an incremental win. Omni-reference is the part with no direct equivalent in a competing model right now.

Every video model on the market takes text, and most take images. A few take audio and video. Wan 3.0 takes a PowerPoint file. It takes an Excel sheet. It takes a PDF and a Keynote and a webpage, and treats their contents as creative reference material rather than as a prompt you have to write yourself.

Consider what that removes from a workflow. Turning a quarterly report into a video currently means reading the report, deciding what matters, writing a prompt that describes it, generating, and then checking whether the output actually reflects the source. Omni-reference collapses the middle of that chain. The document is the input.

The open question — and it is a real one, not a rhetorical one — is what the model actually extracts. A slide deck contains an argument, a visual design, a data set, and a narrative order. Whether Wan 3.0 reads all four, or mostly parses text and ignores layout, will determine whether this is a workflow change or a demo feature. Nothing in the launch announcement answers that, and it is the first thing worth testing.

Alibaba names document-to-video creation explicitly as a target scenario alongside advertising, e-commerce, filmmaking, character animation, and video editing, which suggests the capability is meant to be load-bearing rather than decorative.

One practical caveat: document reference is a launch-day capability of the Alibaba model, and third-party platforms tend to expose reference types in stages. Image, video, and audio references usually arrive first; document parsing may follow in a later integration pass. Check which reference types your access route actually accepts before planning a workflow around spreadsheets.

What Public Beta Status Actually Means

The Wan 3.0 release date marks the start of a beta, not a general availability launch, and the distinction has practical consequences for anyone planning to build on it.

Alibaba’s own wording is that full API access is rolling out soon. That means access today is not the same as access next month, and it is not a stable base for a production integration. Three things commonly change between beta and GA in this product line: the model identifier string, the parameter names, and the pricing.

Identifier drift is the one that bites hardest. Alibaba versions its model IDs by snapshot date — the Wan 2.7 image-to-video documentation uses wan2.7-i2v-2026-04-25 rather than a bare wan2.7-i2v — and a stale identifier returns a plain 404 with no useful error message. The current wan3.0-video string is very likely to acquire a dated suffix at general availability.

Practical guidance: evaluate now, prototype now, but do not ship a customer-facing pipeline against a beta identifier without a fallback path.

How This Lands Against the Competition

As of the Wan 3.0 release date, the 30-second single-pass ceiling puts the model at the top of the published duration range among major hosted video models, most of which sit between 5 and 15 seconds per generation. That advantage is real but narrow, since duration is the easiest specification for a competitor to match in a subsequent release.

Omni-reference is the harder thing for a competitor to copy quickly, because it requires a document-understanding pipeline feeding a video model rather than just a longer context window.

Resolution is where Wan 3.0 does not lead. Holding at 1080p while charging $0.20 per second means a 30-second clip costs about six dollars at the maximum quality the model offers. Whether that is competitive depends entirely on output quality at second 30, which no specification sheet can tell you.

Who This Actually Changes Things For

Anyone doing document-to-video work. This is the standout use case of the Wan 3.0 release date. Feeding a deck or spreadsheet directly as a reference removes an entire manual step from report-to-video and product-sheet-to-ad workflows. No competing model currently advertises this.

Short-form creators. The most immediately usable thing about the Wan 3.0 release date is that thirty seconds in one pass covers a full Reels, Shorts, or TikTok unit without stitching, which is where continuity errors usually appear.

Advertisers and e-commerce teams. Alibaba names advertising, e-commerce, and character animation as target scenarios, and the unified reference-plus-editing workflow means product consistency and revisions happen in the same model.

Developers. The model identifier exists now, so you can build against it — but with full API access still rolling out, hold off on production commitments until Alibaba confirms general availability.

Local-inference users. The Wan 3.0 release date changed nothing for you. No weights, no license, no repository accompanied the launch, consistent with Wan 2.5, 2.6, and 2.7. Alibaba has not open-sourced a Wan video model since Wan 2.2 in July 2025.

Will Wan 3.0 Be Open Sourced?

Nothing about the Wan 3.0 release date or the launch materials suggests it. The public beta is hosted-only across Model Studio and Qwen Cloud, with per-second commercial pricing and no accompanying model card or checkpoint release.

The precedent is not encouraging. Wan 2.1 and Wan 2.2 were Apache 2.0. Wan 2.5 was presented as an open release and the weights never appeared. Wan 2.6 shipped closed. Wan 2.7 shipped closed. Wan 3.0 makes four consecutive hosted-only releases, and two GitHub issues asking about Wan 2.5 open-sourcing have sat unresolved since September 2025.

A useful rule from the Wan 2.7 cycle applies here too: if weights are not downloadable within 30 days of launch, treat any open-source expectation as aspirational.

Generate video with Wan 2.7 on jxp.com today

The Wan 3.0 Release Date in Context: How the Version Line Got Here

Reading the Wan 3.0 release date against the version history makes the trajectory legible.

• Wan 2.1 — February 2025, Apache 2.0, open weights

• Wan 2.2 — July 28, 2025, Apache 2.0. A 27B Mixture-of-Experts diffusion transformer with 14B active parameters, plus a 5B variant that runs on an RTX 4090. The last open release

• Wan 2.5-Preview — announced at Alibaba’s Apsara Conference, September 24, 2025. Native audio-video sync, 10-second clips, 1080P at 24fps

• Wan 2.6 — December 16, 2025. Duration to 15 seconds, role-play and storyboard control, multi-shot switching

• Wan 2.7 — image models April 1, 2026; video suite the following week. Four separate models, 2–15 seconds, 720P and 1080P

• Wan 3.0 — August 6, 2026. Unified model, 30 seconds, omni-reference

Two patterns stand out. Duration roughly doubled at each of the last three steps: 5 to 10 to 15 to 30 seconds. Resolution has been frozen at 1080p since Wan 2.5-Preview in September 2025, across four consecutive releases and nearly a year. Anyone extrapolating a 4K jump at the Wan 3.0 release date was arguing against eleven months of evidence.

The release cadence has also been roughly four months between major versions — September, December, April, August — which is the most defensible basis for guessing when a successor might appear.

What to Test First

Now that the Wan 3.0 release date has passed and access is open, the questions worth answering are empirical rather than speculative. Specification sheets stop being useful at this point; outputs take over.

What still looks consistent at second 30? Duration is the headline, but longer clips give temporal errors more room to accumulate. Character identity, wardrobe, props, spatial geometry, camera continuity and lighting all get harder to hold as a sequence grows. Generate a full 30 seconds with a single character and check frame 1 against frame 720.

How literal is document parsing? Omni-reference is new enough that nobody knows whether it extracts semantic content, visual layout, or both. Feed it a slide deck and see whether the output reflects the argument or the design.

Does 480p cost efficiency hold up? At $0.05 per second, 480p is a quarter the price of 1080p. If it is usable for previz and iteration, that changes the economics of an entire production loop.

Does the unified model hold up on editing? Wan 2.7’s dedicated editing model was purpose-built. Whether a unified model matches it on editing specifically is an open question that no announcement can answer.

Frequently Asked Questions

What is the Wan 3.0 release date?

The Wan 3.0 release date is August 6, 2026, when Alibaba’s Tongyi Lab announced public beta. The model is live now under the identifier wan3.0-video.

Is Wan 3.0 available now?

Yes. Since the Wan 3.0 release date it has been in public beta on Alibaba Cloud Model Studio (international site, ap-southeast-1) and Qwen Cloud. A members-only route on wan.video is described as coming very soon, and full API access is still rolling out.

How much does Wan 3.0 cost?

Published rates are $0.05 per second at 480p, $0.10 per second at 720p, and $0.20 per second at 1080p, in USD. A full 30-second 1080p clip works out to about $6.00.

Does Wan 3.0 support 4K video?

No. The published pricing tops out at 1080p with no 4K tier. The widely repeated 4K claim appears to have originated from the Wan 2.7 image model, where 4096×4096 output is documented for text-to-image tasks only.

How long can Wan 3.0 videos be?

Native 30-second generation in a single pass, without stitching. That is double the 15-second ceiling of Wan 2.6 and Wan 2.7.

What is omni-reference in Wan 3.0?

An input system that accepts structured documents as creative references alongside text, images, audio, and video. Supported formats are doc, xls, ppt, pdf, txt, key, and pages.

Is Wan 3.0 open source?

No weights, license, or repository accompanied the launch. It is the fourth consecutive hosted-only Wan release; Alibaba has not open-sourced a Wan video model since Wan 2.2 in July 2025.

What is the difference between Wan 3.0 and Wan 2.7?

Wan 3.0 doubles clip length to 30 seconds, adds document and spreadsheet reference inputs, and unifies into one model what Wan 2.7 split across four. Maximum resolution is unchanged at 1080p.

Which Wan 3.0 release date predictions were correct?

A Hugging Face community blog predicted August 6, 2026, and was right. Claims of an early or late April 2026 launch, an August 10 launch, and a September 10 launch were all wrong.

Browse the full Wan model lineup on jxp.com